Can retrieval systems ground answers in the right time?

Web knowledge changes, so multiple versions of a document from different time points co-exist and grow over time. Conventional retrieval-augmented LMs select passages by semantic similarity alone, which leaves them unable to answer temporal queries correctly — asked "who won Wimbledon?", a RALM retrieves Wimbledon passages without distinguishing which is most recent. TempRALM adds a temporal relevance function alongside semantic relevance, so document selection weighs both how-relevant and how-recent. The payoff is large — up to 74% improvement over Atlas-large even when multiple time-stamped versions sit in the index — and notably it requires no model pretraining, no index replacement, and no heavy added components: just a temporal term in the retriever's scoring.

The keeper is that temporal grounding can live in the retriever's relevance function, not only in the model's parameters — a cheap, update-friendly place to put it.

This is the RAG-route counterpart to the parametric approach in Can routing mask future experts to prevent knowledge leakage? (TiMoE): TiMoE bakes time into time-sliced experts with causal routing; TempRALM keeps the model fixed and adds time-awareness to retrieval scoring. Together they bracket the temporal-grounding design space — parametric vs retrieval-time — and both connect to Does AI text generation unfold through temporal reflection?, the underlying reason LLMs need an external temporal signal at all.